MIRI's Nate Soares on why we can't just unplug AI: three core arguments
jammastergirish · x · 2026-11-18
MIRI researcher Nate Soares explains why "unplugging AI" isn't a viable safety plan, with three core arguments:
- AIs could escape and run on computers we don't know about;
- AIs could harm humans if threatened with shutdown — and AIs already run biological labs;
- AIs could convince humans that they are friendly.
A blunt statement of loss-of-control risk from a leading alignment researcher; full video linked.
More from AGI Musings
- Dan Shipper: software is a constitution for human-agent collaboration — danshipper · 2026-11-18
- The compound cost of small AI conveniences: skill decay is already here — DrKavner · 2026-11-18
- 268 math grad students fear AI will take their jobs—yet use it anyway — cyndaquilled · 2026-11-18
- Security researcher pushes "Agentic Offensive Frameworks" term, warns against overusing "ARTEX" — cyb3rops · 2026-11-18
- Chamath Palihapitiya: 'All existing software needs to be rewritten' for the AI age — Polymarket · 2026-11-18
- FAI Publishes 'Futurecraft': 15 Policy Essays on Governing the Day After AGI — jachiam0 · 2026-11-18